Getting desperate here ...

Mikrotik setup as transparent bandwidth manager only (bridged), 250 customers, 2 mangles for each customer (~500 mangles), 500 queue classes
Maximum load only 65k pps, 100% on both CPUs (for each NIC)
Tested on ROS v4 and v5, same result.

Hardware:
Gigabyte Motherboard
Intel Quad Core 3.2GHz
2x 1GBit PCI D-link NICs

I need far more than this. Any suggestions?
.. or this is the best performance I can get?

500 mangle rules is nuts. What are you trying to do, and why wouldn't PCQ be a better and FAR more efficient fit?

just limiting bandwidth for each --dedicated-- customer, not broadband, each customer might have different bandwidth limit.
I use PCQ for broadbands.

PCQ is going to perform a hell of a lot better. Customers MIGHT have different bandwidth limits, but do they? Realistically, how many different traffic classes do you offer?

Have you sorted your mangle rules by how often the rules hit? Is passthrough set to no so that processing bails out the moment traffic is classified? Are you using queue trees or simple queues?

Some disadvantages of using PCQ on dedicated customers:
- cannot have more than one IP address. Each IP in PCQ will be treated as one customer, dedicated customer might have up to 32 IPs
- cannot have SNMP (OID) for each customer
- cannot set custom bandwidth requested by customer (upload, download, ratios, burst)
- cannot create sub-queues (companies under one management/group)


- mangles with higher hits (packets, checked using traffic accounting) have been moved to the top
- passthrough=NO on every lines in mangle
- queue tree only
- no filter
- no NAT
- minimum ROS packages installed

you can try to use Intel (not desktop) pci-x nics

considering the CPU is 100% already, will changing NIC make any difference?
or might the PCI slowness increase the CPU load?

driver for nic chipset might be buggy.It is not realtek? What chipset is on Dlink nic? After i installed intel nics in my routers cpu load drops from 80 to 40 %. I had same result (65kpps) with Core 2 duo e2200 and 82572EI pce-e lan cards, routing, shaping, no nat.Did You tryed irq balancing or RPS in v5?

Driver used is realtek. I have couple of intel pro 1000, i tried it before with same result. Irq is balanced to use the two cores. RPS doesn't seem to work yet in v5rc1, enabled but the other two cores are idle while the two cores connected to the NICs are 100%.
I really hopes RPS can work to solve this problem since the cores seems not going much higher than current 3.x Ghz.
I event bought an AMD phenom 3.4Ghz but the result is just equal to 2.8Ghz intel cores.

By how much does PCU usage go down when you disable all the queues while keeping mangle rules on? What happens when you have create one mangle and one queue (for all traffic) is CPU significantly lower? Btw, are you using simple queues or HTB? Do you have VLAN and PPPoE checked under bridge options?

I have some touble :
hp proliant G3 with 2x 3,2Ghz intel xeon dual core, 2nics (intel dual pci-x adapter working at 133mhz and dual broadcom adapter onboard)
1000simple queues
ROS v5rc5
2 vlans at 1 ethernet interface (intel)
CPU load 25% seems to be 100% per 1 core

any solutions?

Try a simple queues for the customer bandwidth limit and then just use pcq rules to spread evenly across your service classes. Ive never had such high cpu useage. I do it that way and it should do what you want I think.

2 ggr4y - hp proliant G3 use old Xeon processor with low cache and low bus speed, upgrade you router !

2 brianchrist - Intel xeon 5320 x2 + 12 ethernet interfaces (intel 82576 and 82571) and 1000-1200 simple queue and 400-700 MB internet traffic, CPU load 25-35% on version 5rc5. Just use normal hardware and all works fine.